Efficient Ad-level Impression Forecasting based on Monotonicity and Sampling

Zhengtao Wu, Lan Zhang, Hui Sheng

Abstract: In the online advertising business, it is important for an advertisement with a specific bid to forecast the number of impressions. Given the reliably predicted impressions, advertisers can set their bids reasonably to guarantee the return on investment. Ad impression forecasting also plays a role in various services, such as bid simulator, high-value advertising screening, and programmatic advertising bid adjustment. Although several impression forecasting tools are already widely used in the industry, previous work is pageview-level methods, which have to calculate the winning score for each impression opportunity and are relatively time-consuming. Besides, due to the data flow of the advertising system is too large, sampling storage leads to loW advertising coverage and inaccurate estimation of winning rate. To address these issues, in this work, we propose a novel adlevel combined method for impression forecasting framework, which combines XGBoost and gaussian distribution modeling based on RANSAC. In particular, XGBoost improves the model’s generalization ability for the missing feature values of ads, and gaussian distribution modeling solves the problems of outlier and sparse sampling While ensuring monotonicity. We deployed the proposed framework in the Tencent online display advertising system and evaluated the effectiveness and efficiency of the framework on a real-world dataset. The experimental results show our design outperforms the existing methods in the accuracy of ad impression forecasting. Meanwhile, our framework achieves 100% ad coverage rate and significantly reduces the average execution time for online serving.